Monday, 26 October 2026

“We should be more ambitious with the software we build”

A new essay on danluu.com starts from Michael Malis (of pgrust) showing off things built with coding agents: a regex engine tunable to beat general-purpose libraries, a ripgrep fork using that engine’s native-code compiler, and a faster Rust build setup, each in “low single-digit hours.” A GPT-6 Astra zstd implementation passed every registered test over 99% of the time at roughly $1 a go.

The honest bit is the caveat: agent loops flounder without clear pass/fail criteria, and the gap between what a model can implement and what it can verify without handholding keeps widening. The test suite becomes the real product.

danluu.com · HN discussion

Five copies of a 27B model, one shared folder, Fable-level scores

GVS5H is a training-free harness: several fresh instances of the same model split a problem and coordinate only through a shared filesystem holding a plan, notes and the current solution. On 100 hard LiveCodeBench v6 problems, orchestrated Qwen3.8 Flash Next hit 93.0% pass@1 against Claude Fable 5’s 90.4%, at about 9% of the cost; a local Qwen3.8-27B went from 66.8% to 92.4%.

The authors say the gains are “not universal” — some of the 11 models tested didn’t improve or got worse. MIT-licensed code.

GVS5H on GitHub · HN discussion

A 2015 image classifier beats frontier LLMs — and the errors tell the story

Lynch Lab fine-tuned Google’s decade-old Inception v3 on 68,011 skin-lesion images and tested it against Claude Fable 5.1 and GPT-6 Astra on 200 biopsy-confirmed cases: 62% first-choice accuracy versus 42% for both LLMs.

The useful part is the failure analysis. At its preset threshold the CNN sent home 18 of 30 melanomas; matched to the LLMs’ operating point, only one or two. Headline accuracy hides where a system fails. Caveats are stated up front: small sample, one run per LLM, public data that may be in training sets.

Lynch Lab · HN discussion

Steal My Bitcoin: a standing invitation to break in

A Show HN puts a wallet key (“about $500”) on a deliberately vulnerable PHP server, readable via a print-flag binary. First to drain the wallet wins. All traffic is logged, and a counter shows the hours since the last successful hack — failure made public as a running scoreboard.

stealmybitcoin.com · HN discussion

A volunteer hub to fix OpenAI’s maths write-ups; Borcherds stops grading homework

A guest post on Terry Tao’s blog by Abhinav Nellore, Johannes Schmitt, Soledad Villar and Rachel Ward launches oai-math-revised, a GitHub project to rewrite OpenAI’s published proofs, which they say “can be difficult to understand and can have incomplete citations.” Commenters ask why volunteers should polish a company’s output, and one flags overstated Unique Games claims.

Separately, Fields medallist Richard Borcherds told The Daily Californian he now grades Berkeley homework on completion only, because AI-assisted students were all scoring top marks.

Terry Tao · Daily Californian

Why AI prose is hard to read, measured

Thomas Hikaru Clark, Matias B. Martin and Kyle Mahowald compared nine frontier models with human writing. All nine used more inanimate subjects with animate verbs (“the analysis reveals”) and more rare n-grams; most also had longer syntactic dependencies and stacked nouns.

Their hypothesis: training for maths, code and reasoning rewards information density over communication. A quick fine-tune with a length penalty made output less human-like.

arXiv · HN discussion

A 1505 knight’s tool book: drills with swappable bits, a chair on casters

The Löffelholz Codex is a 154-page Nuremberg house book by Martin Löffelholz, now in Kraków: interchangeable drill bits, a vise workbench, a swivel chair on castors, a crossbow drawn as an exploded view, a mercury-weighted cradle. A new site is building the first complete transcription and interpretation — 127 pages reviewed, unrevised pages clearly marked as raw AI drafts — with a “Mystery of the Month” series for the drawings nobody can explain yet.

Löffelholz Codex · HN discussion

How Cerebras runs a 3-trillion-parameter model across 59 wafers

Using Kimi K3 (4-bit weights, 8-bit KV cache) as the worked example, Cerebras pipelines the model over 59 WSE-3T wafers with 44 GB each — one wafer per dense layer, two linear-attention layers per wafer. Weights never move; only activations hop between wafers, about 2 µs each, for 1,000–2,000+ tokens a second. A nice dataflow-over-memory design, explained plainly.

Cerebras · HN discussion

Sabre Wulf remastered for the C64, all 256 screens

Genesis Project released its unofficial remake of Ultimate’s 1984 jungle maze, with new graphics by Jon Eggelton, new music and a full map of the 256-screen world, in a 192 KB D64 image for PAL machines. It’s been in previews since 2022 — a long-haul labour of love on a 40-year-old constraint set.

Genesis Project · HN discussion

Antonelli on pole in Austin; Piastri 7th, Norris lap-deleted to 8th

Kimi Antonelli took pole in 1:34.035, ahead of Verstappen, Hadjar and Leclerc. His title rivals start behind: Hamilton 5th, Russell 6th (no engine penalty after all), Piastri 7th, Norris 8th after losing his best lap to track limits — he called the McLaren “easily the hardest car I’ve driven in my life.” Bortoleto takes a 60-place penalty and Stroll starts from the pit lane.

Antonelli clinches today with a win if Russell finishes 9th or lower and Hamilton 6th or lower. The race starts 6 am Monday Sydney time, after this page was written.

Crash.net results · PlanetF1 on Norris · Starting grid

Nakamura’s seventh straight Bullet Brawl; Monte-Carlo retain the women’s Club Cup

Hikaru Nakamura scored 178 against 218 titled players, decided by a 23-game win streak, for his 72nd career Bullet Brawl win (23 this year). Sixteen-year-old Usov opened 17.5/18 before fading to fourth.

At the European Club Cup, Monte-Carlo beat Davit Jojua’s club 4–0 in the last round to take the women’s title on tiebreak over Gambit Bonnevoie — their ninth. Goryachkina and Humpy Koneru won board golds. No new Carlsen news this weekend.

Chess.com · ChessBase

Ex-Klaxons James Righton scores ITV’s Adultery

Righton (Benjamin, Daddy Issues) is scoring Danny Brocklehurst’s six-part drama starring Dominic Cooper and Romola Garai, about two married parents whose affair begins after their teenage children start dating. A quiet weekend for score news otherwise: no John Williams items.

Film Music Reporter

A balloon telescope 35 km up finds a magnetic web above the Sun

Sunrise-III carried a 1 m solar telescope into the stratosphere on a balloon. Analysing its 2024 flight data, Masahito Kubo’s NAOJ team found the quiet Sun’s lower atmosphere isn’t a smooth magnetic canopy but a mesh of thin threads, apparently twisted up by surface motions. They also measured fields inside spicules, the narrow jets of solar material. A reminder that balloons still do frontier science.

ScienceDaily

Tiny snakes that roll go four times faster than they slither

The dwarf reedsnake Pseudorabdion saravacense curls into a wheel and rolls at 1.72–3.69 m/s, about four times its best slithering speed, while staying on a 20 cm track. The Toho University team also worked out how the snakes keep their heads stable mid-roll. Good fuel for anyone designing odd locomotion in a game.

ScienceDaily

Youth Jersey Boys at Hornsby this week; a musical-writers’ incubator at ATYP

Sydney Youth Musical Theatre (performers aged 10–25) is running Jersey Boys at the Hornsby RSL Showroom until Sunday 1 November; tickets are on sale. Their next show is Mary Poppins in July 2027, with auditions opening in December.

ATYP has also announced an Australian Musical Theatre Incubator, 4–8 January, run by Carmel Dean and Maryann Wright, with three bursaries for writers under 26 (applications close 16 November). Note: a new review of We Are The Tigers (Hayes, to 8 Nov) flags teen sex, drugs and horror themes.

SYMT · BroadwayWorld: ATYP · Tigers review

Week ahead: Olympic Park busy three nights, Darling Harbour packed Friday

Air New Zealand’s first WSI–Auckland flight leaves this morning; Sunday’s opening flights ran without reported problems. Lily Allen plays the Entertainment Centre again tonight.

Hans Zimmer is at Afterpay Arena (formerly Qudos Bank) Wednesday and Thursday, finishing around 10:30 pm, then 5 Seconds of Summer there Friday. Friday also brings Richie Jen at the Entertainment Centre and Electric Nation in Tumbalong Park, 5–11:30 pm. BIGBANG is at Accor Stadium on Saturday (Halloween). No metro trackwork confirmed for the week.

Transport NSW events · Darling Harbour what’s on · Travel Weekly